The National Security Agency (NSA), in partnership with cybersecurity agencies from the UK, Australia, New Zealand, and others, has released new guidance aimed at protecting the integrity of data used in AI systems.

The Cybersecurity Information Sheet (CSI), titled AI Data Security: Best Practices for Securing Data Used to Train & Operate AI Systems, outlines emerging threats and sets out 10 recommendations for mitigating them.

The CSI builds on earlier joint guidance from 2024 and signals growing global urgency around safeguarding AI data instead of allowing systems to operate without scrutiny.

The report identifies three core risks across the AI lifecycle: tampered datasets in the supply chain, deliberately poisoned data intended to manipulate models, and data drift-where changes in data over time reduce performance or create new vulnerabilities.

These threats may erode accuracy and trust in AI systems, particularly in sensitive areas like defence, cybersecurity, and critical infrastructure, where even small failures could have far-reaching consequences.

To reduce these risks, the CSI recommends a layered approach-starting with sourcing data from reliable origins and tracking provenance using digital credentials. It advises encrypting data at every stage, verifying integrity with cryptographic tools, and storing data securely in certified systems.

Additional measures include deploying zero trust architecture, using digital signatures for dataset updates, and applying access controls based on data classification instead of relying on broad administrative trust.

The CSI also urges ongoing risk assessments using frameworks like NIST's AI RMF, encouraging organisations to anticipate emerging challenges such as quantum threats and advanced data manipulation.

Privacy-preserving techniques, secure deletion protocols, and infrastructure controls round out the recommendations.

Rather than treating AI as a standalone tool, the guidance calls for embedding strong data governance and security throughout its lifecycle to prevent compromised systems from shaping critical outcomes.